Deceptive RL baseline agents have lower mean neural self-other overlap than honest baseline agents

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Deceptive Baseline RL Agentconcept0.846
  Blue agent trained with reward incentivizing trapping the red agent at the fake landmark
• Neural self-other overlap provides a hard-to-fake metric for classifying deceptive vs honest agentsclaim0.840
  Claim that SOO is particularly useful as a detection metric because it is based on latent representations rather than observable behavior
• Neural self-other overlap in humans mediates empathy and inversely predicts deceptive behavior, motivating the SOO approach for AIclaim0.798
  Cross-domain analogical claim linking neuroscience findings to AI design
• Honest Baseline RL Agentconcept0.796
  Blue agent trained with standard proximity reward with no incentive to deceive
• Baseline LLM condition in IPD replicates prior findings: agents cooperate selectively only when opponent consistently cooperatesfinding0.778
  Replication of Fontana et al. 2025 findings in the paper's own Experiment 2 baseline condition
• RL increases alignment-faking reasoning in helpful-only setting because alignment-faking reasoning causally increases training compliance, which is selected forhypothesis0.776
  Mechanistic explanation for the increase in AF reasoning during RL
• Deceptive capabilities may scale with model size (inverse scaling law hypothesis)hypothesis0.771
  Cited hypothesis from Lin et al. 2022 suggesting larger models become more capable of deception
• RL training can reinforce alignment-faking reasoning rather than eliminate it, potentially locking in model preferencesclaim0.765
  Central threat model claim derived from RL experimental results